Make your agent usage last

Why usage runs out

Every message sends the agent the conversation so far, plus whatever it reads and runs. Claude Code’s cost guide explains that in a long conversation, “a one-line question in a session that has been open all day still draws usage for the whole conversation.” Higher effort, faster speeds, and extra agents working in parallel all add more on top.

Habits that help

• Check before a big job. In the computer app, open the agent settings from the button in the message box that shows your agent’s name. Under the selected agent, each limit your provider reports shows how much is left, such as “40% left.” Point at a limit to see when it resets. Some accounts also show Credits. OpenCode doesn’t report usage, so it shows Usage unavailable. The web app shows Usage unavailable for every agent.
• Start a new conversation for each new task. It’s the biggest saver. See Use one conversation per task.
• Match effort to the task. Use a lower effort for small, clear changes. See Match the effort level to the task.
• Pick a lighter model for routine work. Claude Code’s guide notes that Sonnet “handles most coding tasks well and costs less than Opus.” Codex suggests Luna “for clear, repeatable tasks.” Change it in the agent settings’ Model column.
• Leave Speed on Standard. For Claude Code and Codex, Fast in the agent settings’ Speed column answers sooner, but Enjoy notes that it uses paid credits or costs more.
• Be specific, and plan big changes. Claude Code’s guide says vague requests “trigger broad scanning,” and planning first prevents “expensive re-work.” See Describe what you want and Plan before you build.
• Count your parallel work. Each conversation that runs at the same time uses your plan, and so does each scheduled workflow run.

When you hit a limit

The conversation stops with The usage limit was reached. or The request limit was reached. Your messages and progress are kept. You can:

• Wait for the reset, then select Retry. In the computer app, point at the limit in the agent settings to see when it resets.
• Use another account. In the computer app, Claude Code, Codex, and Grok Build can keep more than one signed-in account. Pick another one from the account button above the usage in the agent settings, then select Retry. Each project remembers which account each agent uses. See Switch between agent accounts.
• Continue with another agent. Choose a different agent in the agent settings. A Retry with button, followed by the agent’s name, appears beside Retry and continues with it. The other agent receives the conversation so far.

An example

You’re polishing a website on Claude Code with Max effort, all in one conversation that started on Monday. By Wednesday the agent settings show 8% left. Next time, you start a new conversation for each page, set effort to Medium for layout tweaks, and keep High for the checkout bug.

Common pitfalls

• Keeping one conversation going for days. The whole history rides along with every message.
• Running several agents on the same idea. You pay for each one. Run agents in parallel on separate tasks. See Run agents in parallel.
• Scheduling a workflow before it’s reliable. Each run uses your plan, even when nobody reads the result. See Turn repeat work into workflows.